def __init_model(self):
input_text = "你好, 很高兴认识你"
self.processor = AutoProcessor.from_pretrained(
pretrained_path=config.baker_mapper_pretrained_path)
input_ids = self.processor.text_to_sequence(input_text, inference=True)
# tacotron2_config = AutoConfig.from_pretrained( config.tacotron2_baker )
# self.tacotron2 = TFAutoModel.from_pretrained( config=tacotron2_config, pretrained_path=config.tacotron2_pretrained_path, training=False, name="tacotron2" )
tacotron2_config = AutoConfig.from_pretrained(config.tacotron2_baker)
self.tacotron2 = TFAutoModel.from_pretrained(
config=tacotron2_config,
pretrained_path=None,
is_build=
False, # don't build model if you want to save it to pb. (TF related bug)
name="tacotron2")
self.tacotron2.setup_window(win_front=5, win_back=5)
_, mel_outputs, stop_token_prediction, alignment_history = self.tacotron2.inference(
tf.expand_dims(tf.convert_to_tensor(input_ids, dtype=tf.int32), 0),
tf.convert_to_tensor([len(input_ids)], tf.int32),
tf.convert_to_tensor([0], dtype=tf.int32))
self.tacotron2.load_weights(config.tacotron2_pretrained_path)
tf.saved_model.save(self.tacotron2,
"./test_saved",
signatures=self.tacotron2.inference)
self.tacotron2 = tf.saved_model.load("./test_saved")
mb_melgan_config = AutoConfig.from_pretrained(
config.multiband_melgan_baker)
self.mb_melgan = TFAutoModel.from_pretrained(
config=mb_melgan_config,
pretrained_path=config.multiband_melgan_pretrained_path,
name="mb_melgan")
self.processor = AutoProcessor.from_pretrained(
pretrained_path=config.baker_mapper_pretrained_path)
def __init__(self):
self.converter = opencc.OpenCC('tw2s.json')
tts_model_dir = os.environ['TTS_MODEL_DIR']
if not os.path.exists(tts_model_dir):
parent_dir = os.path.dirname(tts_model_dir)
zip_file_path = os.path.join(parent_dir, 'model_files.zip')
download_blob('dailybrief', 'models/model_files.zip', zip_file_path)
with zipfile.ZipFile(zip_file_path, 'r') as zip_ref:
zip_ref.extractall(parent_dir)
tacotron2_config_file = os.path.join(
tts_model_dir, 'config/tacotron2.baker.v1.yaml')
mb_melgan_config_file = os.path.join(
tts_model_dir, 'config/multiband_melgan.baker.v1.yaml')
tacotron2_config = AutoConfig.from_pretrained(tacotron2_config_file)
mb_melgan_config = AutoConfig.from_pretrained(mb_melgan_config_file)
text2mel_model_file = os.path.join(
tts_model_dir, 'models/tacotron-model-100000.h5')
vocoder_model_file = os.path.join(tts_model_dir, 'models/generator-920000.h5')
baker_mapper_file = os.path.join(tts_model_dir, 'models/baker_mapper.json')
self.text2mel_model = TFAutoModel.from_pretrained(
config=tacotron2_config,
pretrained_path=text2mel_model_file,
name='tacotron2'
)
self.vocoder_model = TFAutoModel.from_pretrained(
config=mb_melgan_config,
pretrained_path=vocoder_model_file,
name='mb_melgan'
)
self.processor = AutoProcessor.from_pretrained(pretrained_path=baker_mapper_file)
def __init__(self):
if (Text2SpeechModel == "dc_tts"):
self.g = Graph(mode="synthesize")
print("Text2Speech Tensorflow Graph loaded")
elif (Text2SpeechModel == "RTVC"):
enc_model_fpath = os.path.join(
root_file_path, "RTVC", "encoder/saved_models/pretrained.pt")
syn_model_dir = os.path.join(
root_file_path, "RTVC",
"synthesizer/saved_models/logs-pretrained")
voc_model_fpath = os.path.join(
root_file_path, "RTVC",
"vocoder/saved_models/pretrained/pretrained.pt")
encoder.load_model(enc_model_fpath)
self.synthesizer = Synthesizer(os.path.join(
syn_model_dir, "taco_pretrained"),
low_mem=False)
vocoder.load_model(voc_model_fpath)
in_fpath = os.path.join("/",
*root_file_path.split("/")[:-1],
"REF/refaudioRTVC/ref.wav")
preprocessed_wav = encoder.preprocess_wav(in_fpath)
original_wav, sampling_rate = librosa.load(in_fpath)
preprocessed_wav = encoder.preprocess_wav(original_wav,
sampling_rate)
embed = encoder.embed_utterance(preprocessed_wav)
self.embeds = [embed]
elif (Text2SpeechModel == "AudioSynth"):
taco_pretrained_config_path = os.path.join(
root_file_path,
'AudioSynth/TensorFlowTTS/examples/tacotron2/conf/tacotron2.v1.yaml'
)
tacotron2_config = AutoConfig.from_pretrained(
taco_pretrained_config_path)
taco_path = os.path.join(root_file_path,
"AudioSynth/tacotron2-120k.h5")
self.tacotron2 = TFAutoModel.from_pretrained(
config=tacotron2_config,
pretrained_path=taco_path,
training=False,
name="tacotron2")
melgan_stft_pretrained_config_path = os.path.join(
root_file_path,
'AudioSynth/TensorFlowTTS/examples/melgan.stft/conf/melgan.stft.v1.yaml'
)
melgan_stft_config = AutoConfig.from_pretrained(
melgan_stft_pretrained_config_path)
melgan_stft_path = os.path.join(root_file_path,
"AudioSynth/melgan.stft-2M.h5")
self.melgan_stft = TFAutoModel.from_pretrained(
config=melgan_stft_config,
pretrained_path=melgan_stft_path,
name="melgan_stft")
self.processor = AutoProcessor.from_pretrained(
pretrained_path=os.path.join(
root_file_path, "AudioSynth/ljspeech_mapper.json"))
mels, alignment_history, audios = do_synthesis(
"Hello, how can I help you today?", self.tacotron2,
self.melgan_stft, "TACOTRON", "MELGAN-STFT", self.processor)
def init_speech():
global sf
global tf
global TFAutoModel
global AutoProcessor
import soundfile as sf
import tensorflow as tf
from tensorflow_tts.inference import TFAutoModel
from tensorflow_tts.inference import AutoProcessor
global fastspeech2
global mb_melgan
global processor
# initialize fastspeech2 model.
fastspeech2 = TFAutoModel.from_pretrained(
"tensorspeech/tts-fastspeech2-ljspeech-en")
# initialize mb_melgan model
mb_melgan = TFAutoModel.from_pretrained(
"tensorspeech/tts-mb_melgan-ljspeech-en")
# inference
processor = AutoProcessor.from_pretrained(
"tensorspeech/tts-fastspeech2-ljspeech-en")
inference("Hello sir")
debug("Speech", "init")
def __init__(self):
# initialize tts model. fastspeech2 or tacotron2
self.tts_model = TFAutoModel.from_pretrained(
"tensorspeech/tts-fastspeech2-kss-ko")
# initialize mb_melgan model
self.mb_melgan = TFAutoModel.from_pretrained(
"tensorspeech/tts-mb_melgan-kss-ko")
# inference
self.processor = AutoProcessor.from_pretrained(
"tensorspeech/tts-fastspeech2-kss-ko")
def test_auto_model(config_path):
config = AutoConfig.from_pretrained(pretrained_path=config_path)
model = TFAutoModel.from_pretrained(pretrained_path=None, config=config)
# test save_pretrained
config.save_pretrained("./test_saved")
model.save_pretrained("./test_saved")
# test from_pretrained
config = AutoConfig.from_pretrained("./test_saved/config.yml")
model = TFAutoModel.from_pretrained("./test_saved/model.h5", config=config)
def __init_model(self):
tacotron2_config = AutoConfig.from_pretrained(config.tacotron2_baker)
self.tacotron2 = TFAutoModel.from_pretrained(
config=tacotron2_config, pretrained_path=config.tacotron2_pretrained_path, training=False, name="tacotron2")
self.tacotron2.setup_window(win_front=5, win_back=5)
mb_melgan_config = AutoConfig.from_pretrained(config.multiband_melgan_baker)
self.mb_melgan = TFAutoModel.from_pretrained(
config=mb_melgan_config, pretrained_path=config.multiband_melgan_pretrained_path, name="mb_melgan")
self.processor = AutoProcessor.from_pretrained(pretrained_path=config.baker_mapper_pretrained_path)
def __init__(self):
# gpu memory의 1/3 만을 할당하기로 제한
gpu_options = tf.compat.v1.GPUOptions(
per_process_gpu_memory_fraction=0.8)
conf = tf.compat.v1.ConfigProto(gpu_options=gpu_options)
# 탄력적인 메모리 할당
#conf.gpu_options.allow_growth = True
session = tf.compat.v1.Session(config=conf)
# tacotron 설정, 학습된 모델 가져오기
module_path = os.path.dirname(os.path.abspath(__file__))
tacotron2_config = AutoConfig.from_pretrained(
os.path.join(module_path,
'examples/tacotron2/conf/tacotron2.song44k.v5.yaml'))
self.tacotron2 = TFAutoModel.from_pretrained(
config=tacotron2_config,
pretrained_path=os.path.join(
module_path,
"examples/tacotron2/exp/train.tacotron2.song44k.v5/checkpoints/model-68000.h5"
),
name="tacotron2")
# fast speech 설정, 학습된 모델 가져오기
fastspeech2_config = AutoConfig.from_pretrained(
os.path.join(
module_path,
'examples/fastspeech2/conf/fastspeech2.song44k.v5.1.yaml'))
self.fastspeech2 = TFAutoModel.from_pretrained(
config=fastspeech2_config,
pretrained_path=os.path.join(
module_path,
"examples/fastspeech2/exp/train.fastspeech2.song44k.v5.1/checkpoints/model-600000.h5"
),
name="fastspeech2")
# mel gan 설정, 학습된 모델 가져오기
mb_melgan_config = AutoConfig.from_pretrained(
os.path.join(
module_path,
'examples/multiband_melgan/conf/multiband_melgan.ko.song44k.v5.1.yaml'
))
self.mb_melgan = TFAutoModel.from_pretrained(
config=mb_melgan_config,
pretrained_path=os.path.join(
module_path,
"examples/multiband_melgan/exp/train.multiband_melgan.ko.song44k.v5.1/checkpoints/generator-1740000.h5"
),
name="mb_melgan")
#processor - 글자 별 상응하는 숫자의 mapper 설정 가져오기
self.processor = AutoProcessor.from_pretrained(
pretrained_path=os.path.join(module_path,
"test/files/kss_mapper.json"))
def __init__(self):
# gpu memory의 1/3 만을 할당하기로 제한
gpu_options = tf.compat.v1.GPUOptions(per_process_gpu_memory_fraction=0.8)
conf = tf.compat.v1.ConfigProto(gpu_options=gpu_options)
# 탄력적인 메모리 할당
#conf.gpu_options.allow_growth = True
session = tf.compat.v1.Session(config=conf)
# tacotron 설정, 학습된 모델 가져오기
module_path = os.path.dirname(os.path.abspath(__file__))
# pdb.set_trace()
tacotron2_config = AutoConfig.from_pretrained(os.path.join(module_path,'./examples/tacotron2/conf/tacotron2.song8k.v3.yaml'))
self.tacotron2 = TFAutoModel.from_pretrained(
config=tacotron2_config,
pretrained_path=os.path.join(module_path,"./examples/tacotron2/exp/train.tacotron2.song8k.v3/checkpoints/model-68000.h5"),
name="tacotron2"
)
# fast speech 설정, 학습된 모델 가져오기
fastspeech2_config = AutoConfig.from_pretrained(os.path.join(module_path,'./examples/fastspeech2/conf/fastspeech2.song8k.v3.yaml'))
self.fastspeech2 = TFAutoModel.from_pretrained(
config=fastspeech2_config,
# pretrained_path=os.path.join(module_path,"./examples/fastspeech2/exp/train.fastspeech2.song8k.v1.1/checkpoints/model-200000.h5"),
pretrained_path=os.path.join(module_path,"./examples/fastspeech2/exp/train.fastspeech2.song8k.v3/checkpoints/model-200000.h5"),
name="fastspeech2"
)
# fastspeech1_config = AutoConfig.from_pretrained(os.path.join(module_path,'examples/fastspeech/conf/fastspeech.v3_song44k_v51.yaml'))
# self.fastspeech1 = TFAutoModel.from_pretrained(
# config=fastspeech1_config,
# pretrained_path=os.path.join(module_path,"examples/fastspeech/exp/train.fastspeech.song.v41/checkpoints/model-200000.h5"),
# name="fastspeech1"
# )
# resizing positional embedding
# self.fastspeech1._build()
# self.fastspeech1.save_weights("./resize.h5")
# self.fastspeech1.resize_positional_embeddings(8000)
# self.fastspeech1.load_weights("./resize.h5", by_name=True, skip_mismatch=True)
# mel gan 설정, 학습된 모델 가져오기
mb_melgan_config = AutoConfig.from_pretrained(os.path.join(module_path,'./examples/multiband_melgan/conf/multiband_melgan.ko.8k.v3.yaml'))
self.mb_melgan = TFAutoModel.from_pretrained(
config=mb_melgan_config,
pretrained_path=os.path.join(module_path,"./examples/multiband_melgan/exp/train.multiband_melgan.ko.song8k.v3/checkpoints/generator-1000000.h5"),
name="mb_melgan"
)
#processor - 글자 별 상응하는 숫자의 mapper 설정 가져오기
self.processor = AutoProcessor.from_pretrained(pretrained_path=os.path.join(module_path,"test/files/kss_mapper.json"))
def _converter_model(self):
with open( config.tacotron2_baker ) as f:
conf = yaml.load(f, Loader=yaml.Loader)
conf = Tacotron2Config(**conf["tacotron2_params"])
self.tacotron2 = TFTacotron2(config=conf, training=False, name="tacotron2", enable_tflite_convertible=True)
self.tacotron2.setup_window(win_front=5, win_back=5)
self.tacotron2.setup_maximum_iterations(1000) # be careful
self.tacotron2._build()
self.tacotron2.load_weights(config.tacotron2_pretrained_path)
tacotron2_concrete_function = self.tacotron2.inference_tflite.get_concrete_function()
converter = tf.lite.TFLiteConverter.from_concrete_functions( [tacotron2_concrete_function] )
converter.optimizations = [tf.lite.Optimize.DEFAULT]
converter.target_spec.supported_ops = [ tf.lite.OpsSet.TFLITE_BUILTINS, tf.lite.OpsSet.SELECT_TF_OPS ]
tflite_model = converter.convert()
with open('tacotron2.tflite', 'wb') as f:
f.write(tflite_model)
print('Model size is %f MBs.' % (len(tflite_model) / 1024 / 1024.0) )
#tacotron2_config = AutoConfig.from_pretrained( config.tacotron2_baker )
#self.tacotron2 = TFAutoModel.from_pretrained( config=tacotron2_config, pretrained_path='tacotron2.tflite', training=False, name="tacotron2" )
#self.tacotron2.setup_window(win_front=5, win_back=5)
self.interpreter = tf.lite.Interpreter(model_path='tacotron2.tflite')
self.interpreter.allocate_tensors()
self.input_details = self.interpreter.get_input_details()
self.output_details = self.interpreter.get_output_details()
mb_melgan_config = AutoConfig.from_pretrained( config.multiband_melgan_baker )
self.mb_melgan = TFAutoModel.from_pretrained( config=mb_melgan_config, pretrained_path=config.multiband_melgan_pretrained_path, name="mb_melgan" )
self.processor = AutoProcessor.from_pretrained(pretrained_path=config.baker_mapper_pretrained_path)
# The Original Code is Copyright (C) 2020 Voxell Technologies.
# All rights reserved.
import soundfile as sf
import tensorflow as tf
from tensorflow_tts.inference import AutoConfig
from tensorflow_tts.inference import TFAutoModel
from tensorflow_tts.inference import AutoProcessor
processor = AutoProcessor.from_pretrained("../ljspeech_mapper.json")
# initialize tacotron2 model
config = AutoConfig.from_pretrained("../tacotron2/conf/tacotron2.v1.yaml")
tacotron2 = TFAutoModel.from_pretrained(
config=config,
pretrained_path=None,
is_build=True,
name="tacotron2"
)
tacotron2.setup_window(win_front=6, win_back=6)
tacotron2.setup_maximum_iterations(3000)
tacotron2.load_weights("../tacotron2/checkpoints/model-120000.h5")
# tf.saved_model.save(tacotron2, "../tacotron2/inference", signatures=tacotron2.inference)
# tacotron2 = tf.saved_model.load("../tacotron2/inference")
# initialize melgan model
melgan_config = AutoConfig.from_pretrained('../melgan/conf/melgan.v1.yaml')
melgan = TFAutoModel.from_pretrained(
config=melgan_config,
pretrained_path="../melgan/checkpoints/generator-1500000.h5"
import yaml
import numpy as np
import matplotlib.pyplot as plt
import scipy.io.wavfile
import re
from tensorflow_tts.inference import TFAutoModel
from tensorflow_tts.inference import AutoConfig
from tensorflow_tts.inference import AutoProcessor
fastspeech2_config = AutoConfig.from_pretrained('pretrained/fastspeech2_config.yml')
#fastspeech2_config.max_position_embeddings = 20000
fastspeech2 = TFAutoModel.from_pretrained(
config=fastspeech2_config,
pretrained_path="pretrained/fastspeech2-150k.h5",
name="fastspeech2"
)
mb_melgan_config = AutoConfig.from_pretrained('pretrained/mb.melgan_config.yml')
mb_melgan = TFAutoModel.from_pretrained(
config=mb_melgan_config,
pretrained_path="pretrained/mb.melgan-940k.h5",
name="mb_melgan"
)
processor = AutoProcessor.from_pretrained(pretrained_path="pretrained/ljspeech_mapper.json")
def do_synthesis(input_text, text2mel_model, vocoder_model, text2mel_name, vocoder_name):
input_ids = processor.text_to_sequence(input_text)
from tensorflow_tts.configs.tacotron2 import Tacotron2Config
from tensorflow_tts.models.tacotron2 import TFTacotron2
from tensorflow_tts.inference import AutoConfig
from tensorflow_tts.inference import TFAutoModel
from tensorflow_tts.inference import AutoProcessor
starttime = datetime.datetime.now()
# Tacotron2
tacotron2_config = AutoConfig.from_pretrained('examples/tacotron2/conf/tacotron2.baker.v1.yaml')
tacotron2 = TFAutoModel.from_pretrained(
config=tacotron2_config,
# pretrained_path="trained/model-60000.h5",
# pretrained_path="trained/taco_char_server_126k.h5",
pretrained_path="trained/taco_local_word_new_82k.h5",
name="tacotron2"
)
# FastSpeech2
fastspeech2_config = AutoConfig.from_pretrained('examples/fastspeech2/conf/fastspeech2.baker.v2.yaml')
fastspeech2 = TFAutoModel.from_pretrained(
config=fastspeech2_config,
pretrained_path="trained/fs_local_135k.h5",
name="fastspeech2"
)
# MB-MelGAN
mb_melgan_config = AutoConfig.from_pretrained('examples/multiband_melgan/conf/multiband_melgan.baker.v1.yaml')
mb_melgan = TFAutoModel.from_pretrained(
def test_auto_model(config_path):
config = AutoConfig.from_pretrained(pretrained_path=config_path)
model = TFAutoModel.from_pretrained(config=config, pretrained_path=None)
# The Original Code is Copyright (C) 2020 Voxell Technologies.
# All rights reserved.
import soundfile as sf
import tensorflow as tf
from tensorflow_tts.inference import AutoConfig
from tensorflow_tts.inference import TFAutoModel
from tensorflow_tts.inference import AutoProcessor
processor = AutoProcessor.from_pretrained("../ljspeech_mapper.json")
# initialize fastspeech2 model.
config = AutoConfig.from_pretrained('../fastspeech2/conf/fastspeech2.v1.yaml')
fastspeech2 = TFAutoModel.from_pretrained(
config=config,
pretrained_path="../fastspeech2/checkpoints/model-150000.h5",
is_build=True,
name="fastspeech2")
# initialize melgan model
melgan_config = AutoConfig.from_pretrained('../melgan/conf/melgan.v1.yaml')
melgan = TFAutoModel.from_pretrained(
config=melgan_config,
pretrained_path="../melgan/checkpoints/generator-1670000.h5")
input_text = "how much wood would a woodchuck chuck if a woodchuck could chuck wood?"
input_ids = processor.text_to_sequence(input_text)
# fastspeech2 inference
mel_before, mel_after, duration_outputs, _, _ = fastspeech2.inference(
input_ids=tf.expand_dims(tf.convert_to_tensor(input_ids, dtype=tf.int32),
# )
# # FastSpeech2
# fastspeech2_config = AutoConfig.from_pretrained('examples/fastspeech2/conf/fastspeech2.baker.v2.yaml')
# fastspeech2 = TFAutoModel.from_pretrained(
# config=fastspeech2_config,
# pretrained_path="trained/fastspeech2-200k.h5",
# name="fastspeech2"
# )
# MB-MelGAN
mb_melgan_config = AutoConfig.from_pretrained(
'examples/multiband_melgan/conf/multiband_melgan.baker.v1.yaml')
mb_melgan = TFAutoModel.from_pretrained(
config=mb_melgan_config,
pretrained_path=
"trained/mb.melgan_word_428k.h5", # "trained/mb.melgan-1M.h5"
# is_build=False, # don't build model if you want to save it to pb. (TF related bug)
name="mb_melgan")
# LJSpeechProcessor
# processor = AutoProcessor.from_pretrained("./tensorflow_tts/processor/pretrained/ljspeech_mapper.json")
# processor = AutoProcessor.from_pretrained("trained/baker_mapper_mix.json")
processor = AutoProcessor.from_pretrained("trained/baker_mapper_word.json")
# save tacotron2 to pb
# def save_tacotron2_pb():
# input_text = "i love you so much."
# input_ids = processor.text_to_sequence(input_text)
#
# tacotron2.setup_window(win_front=6, win_back=6)
# tacotron2.setup_maximum_iterations(3000)
parser = argparse.ArgumentParser(description='Start TTS and Vocoder')
parser.add_argument(
'--path_fs',
default=
"examples/fastspeech2_libritts/outdir_libri/checkpoints/model-855000.h5"
)
parser.add_argument('--path_mb',
default="checks/mb_melgan_or/mb.melgan-940k.h5")
args = parser.parse_args()
fastspeech2_config = AutoConfig.from_pretrained(
'examples/fastspeech2/conf/fastspeech2.v1.yaml')
fastspeech2 = TFAutoModel.from_pretrained(
config=fastspeech2_config,
pretrained_path=args.
path_fs, #"examples/fastspeech2_libritts/outdir_libri/checkpoints/model-855000.h5",
#training=False,
name="fastspeech2")
mb_melgan_config = AutoConfig.from_pretrained(
'examples/multiband_melgan/conf/multiband_melgan.v1.yaml')
mb_melgan = TFAutoModel.from_pretrained(
config=mb_melgan_config,
pretrained_path=args.
path_mb, #"checks/mb_melgan_or/mb.melgan-940k.h5",
name="mb_melgan")
processor = AutoProcessor.from_pretrained(
pretrained_path="dump_ljspeech/ljspeech_mapper.json")
app.run(host='0.0.0.0', port=5454)
# import IPython.display as ipd
from tensorflow.python.framework.convert_to_constants import convert_variables_to_constants_v2
from tensorflow_tts.configs.tacotron2 import Tacotron2Config
from tensorflow_tts.models.tacotron2 import TFTacotron2
from tensorflow_tts.inference import AutoConfig
from tensorflow_tts.inference import TFAutoModel
from tensorflow_tts.inference import AutoProcessor
starttime = datetime.datetime.now()
# Tacotron2
tacotron2_config = AutoConfig.from_pretrained(
'examples/tacotron2/conf/tacotron2.baker.v1.yaml')
tacotron2 = TFAutoModel.from_pretrained(
config=tacotron2_config,
pretrained_path="trained/taco_server_word_60k.h5",
name="tacotron2")
# # FastSpeech2
# fastspeech2_config = AutoConfig.from_pretrained('examples/fastspeech2/conf/fastspeech2.baker.v2.yaml')
# fastspeech2 = TFAutoModel.from_pretrained(
# config=fastspeech2_config,
# pretrained_path="trained/fastspeech2-200k.h5",
# name="fastspeech2"
# )
# MB-MelGAN
mb_melgan_config = AutoConfig.from_pretrained(
'examples/multiband_melgan/conf/multiband_melgan.baker.v1.yaml')
mb_melgan = TFAutoModel.from_pretrained(
config=mb_melgan_config,
from tensorflow_tts.inference import TFAutoModel
from tensorflow_tts.inference import AutoProcessor
# 这里会下载2个东西
"""
[nltk_data] Downloading package averaged_perceptron_tagger to
[nltk_data] /root/nltk_data...
[nltk_data] Unzipping taggers/averaged_perceptron_tagger.zip.
[nltk_data] Downloading package cmudict to /root/nltk_data...
[nltk_data] Unzipping corpora/cmudict.zip.
"""
tacotron2_config = AutoConfig.from_pretrained(
'TensorFlowTTS/examples/tacotron2/conf/tacotron2.baker.v1.yaml')
tacotron2 = TFAutoModel.from_pretrained(config=tacotron2_config,
pretrained_path="tacotron2-100k.h5",
training=False,
name="tacotron2")
mb_melgan_config = AutoConfig.from_pretrained(
'TensorFlowTTS/examples/multiband_melgan/conf/multiband_melgan.baker.v1.yaml'
)
mb_melgan = TFAutoModel.from_pretrained(config=mb_melgan_config,
pretrained_path="mb.melgan-920k.h5",
name="mb_melgan")
processor = AutoProcessor.from_pretrained(
pretrained_path="./baker_mapper.json")
def do_synthesis(input_text, text2mel_model, vocoder_model, text2mel_name,
vocoder_name):
input_ids = processor.text_to_sequence(input_text, inference=True)
# Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301, USA.
# The Original Code is Copyright (C) 2020 Voxell Technologies.
# All rights reserved.
import soundfile as sf
import tensorflow as tf
from tensorflow_tts.inference import TFAutoModel, AutoConfig, AutoProcessor
processor = AutoProcessor.from_pretrained("../ljspeech_mapper.json")
# initialize fastspeech model.
fs_config = AutoConfig.from_pretrained('../fastspeech/conf/fastspeech.v1.yaml')
fastspeech = TFAutoModel.from_pretrained(
config=fs_config,
pretrained_path="../fastspeech/checkpoints/model-195000.h5")
# initialize melgan model
melgan_config = AutoConfig.from_pretrained('../melgan/conf/melgan.v1.yaml')
melgan = TFAutoModel.from_pretrained(
config=melgan_config,
pretrained_path="../melgan/checkpoints/generator-1670000.h5")
ids = processor.text_to_sequence(
"how much wood would a woodchuck chuck if a woodchuck could chuck wood?")
print(ids)
ids = tf.expand_dims(ids, 0)
# fastspeech inference
masked_mel_before, masked_mel_after, duration_outputs = fastspeech.inference(